Modeling each unit singly and finding a wholly bottom-up solution works, I guess, but I wonder if it's worth coming at it from a tactical level. You're the commander pushing entire divisions across the battlefield, and the individual units in the division would be trained to maneuver together. They're soldiers with a commander, not individual ants. Shouldn't the model reflect that?
I wonder if a better solution would be to treat the whole group of units as one mega-unit, pathfind for that, then if the path for the megaunit is some factor longer than the straight collision-ignoring route, split it up into 2-6 squad-units. Or split the megaunit up into however many squad-units it takes to have a maximum of a dozen units per squad, pathfind for the squads with some collision avoidance, then do a more fine-grained process of calculating a path for each unit from its current position relative to its squad's center to the same position relative to the next point on the squad's path.
I never play this sort of game so maybe this is routine now, I dunno.
I wonder if a better solution would be to treat the whole group of units as one mega-unit, pathfind for that, then if the path for the megaunit is some factor longer than the straight collision-ignoring route, split it up into 2-6 squad-units. Or split the megaunit up into however many squad-units it takes to have a maximum of a dozen units per squad, pathfind for the squads with some collision avoidance, then do a more fine-grained process of calculating a path for each unit from its current position relative to its squad's center to the same position relative to the next point on the squad's path.
I never play this sort of game so maybe this is routine now, I dunno.